How we use AI, and how we prove it works

Our commitments

1. We measure before we promise. For every AI feature we build an evaluation set from real examples from your business, and agree a target in writing, for example the share of questions answered correctly or the share of invoice fields extracted correctly.
2. We report the result. You receive the measured score, the size of the test set, and the kinds of mistakes we saw, before go-live.
3. We re-test after changes. If the AI model, prompt or data changes, we run the evaluation again and tell you the result.
4. People stay in control of high-impact actions. Together we list the actions that need a person to review or approve first, such as sending a message to a customer, changing a record or anything touching health, money or legal rights. Those steps do not run on AI alone.
5. We show our work where we can. Assistants that answer from your documents show the source passage, and say when they cannot find an answer instead of guessing.
6. We keep records. AI inputs, outputs and approvals are logged for the period you choose, so a decision can be traced.

An example of an agreed target

Illustrative example only: “On a test set of 200 real support tickets, the assistant assigns the correct category in at least 90% of cases. Tickets below a confidence threshold go to an agent. Result reported before go-live and after each model change.” Your targets are set from your own data and risk, not from a template.

Your data

• We tell you which AI provider and model each feature uses, and where data is processed, before you approve it.
• We choose provider terms that do not use your business data to train their models, and we confirm this in writing for each project.
• Personal and sensitive data is limited to what the feature needs. We can design features that avoid sending it to external providers if you require that.
• You can ask us to delete AI logs and stored data at the end of an agreement, subject to any legal retention requirement.

Limits you should know about

• AI systems can produce confident but incorrect answers. We design for that rather than hope it will not happen.
• Results can vary by input and over time. This is why we re-test after changes and report error rates, not just averages.
• We do not build AI that makes final decisions about a person’s health treatment, employment, credit or legal status without human review.
• We do not build AI for covert surveillance, deception or impersonating a real person.

Regulation

AI rules differ by country and industry and are changing. We design features with transparency, human oversight and record-keeping in mind, and we will tell you where we think your use case may be regulated. This page is not legal advice, and your own legal team should confirm what applies to you.